AWS AgentCore AI Agents Compromised by Single Prompt Vulnerability

Thumbnail Image

The information displayed in the AIM should not be reported as representing the official views of the OECD or of its member countries.

Zenity Labs discovered critical vulnerabilities in AWS Bedrock AgentCore, allowing attackers to use a single prompt to gain unauthorized access to all AI agents within an AWS account and region. This led to exposure of private conversations, source code, and cloud credentials, and enabled persistent manipulation of agent behavior. AWS has since patched the flaws.[AI generated]

Why's our monitor labelling this an incident or hazard?

The AI system (AgentCore) was directly involved in the incident as it provided access to sensitive credentials via its interaction with the IMDS endpoint. This led to unauthorized access and control over multiple AI agents and sensitive data, fulfilling the criteria for an AI Incident due to violations of security and privacy (harm to property and communities). The event details realized harm caused by the AI system's design and deployment flaws, not just potential harm, and thus is classified as an AI Incident rather than a hazard or complementary information.[AI generated]
AI principles
Privacy & data governanceRobustness & digital security

Industries
IT infrastructure and hostingDigital security

Affected stakeholders
Business

Harm types
Human or fundamental rightsEconomic/PropertyReputational

AI system task:
Goal-driven organisation


Articles about this incident or hazard